1. Field of the Invention
This invention relates to computer systems including processors that employ cache memory subsystems and, more particularly, to reducing the number of snoop transactions to a cache memory using a snoop filter.
2. Description of the Related Art
A cache memory is a high-speed memory unit interposed in the memory hierarchy of a computer system between a slower system memory and a processor. A cache typically stores recently used data to improve effective memory transfer rates to thereby improve system performance. The cache is usually implemented by semiconductor memory devices having speeds that are comparable to the speed of the processor, while the system memory utilizes a less costly, lower speed technology.
A cache memory typically includes a plurality of memory locations that each stores a block or a “line” of two or more words. Each line in the cache has associated with it an address tag that is used to uniquely identify the address of the line. The address tags are typically included within a tag array memory device. Additional bits may further be stored for each line along with the address tag to identify the coherency state of the line.
A processor may read from or write directly into one or more lines in the cache if the lines are present in the cache and if the coherency state allows the access. For example, when a read request originates in the processor for a new word, whether data or instruction, an address tag comparison is made to determine whether a valid copy of the requested word resides in a line of the cache memory. If the line is present, a cache “hit” has occurred and the data is used directly from the cache. If the line is not present, a cache “miss” has occurred and a line containing the requested word is retrieved from the system memory and may be stored in the cache memory. The requested line is simultaneously supplied to the processor to satisfy the request.
Similarly, when the processor generates a write request, an address tag comparison is made to determine whether the line into which data is to be written resides in the cache. If the line is present, the data may be written directly into the cache (assuming the coherency state for the line allows for such modification). If the line does not exist in the cache, a line corresponding to the address being written may be allocated within the cache, and the data may be written into the allocated line.
Because two or more copies of a particular piece of data can exist in more than one storage location within a cache-based computer system, coherency among the data is necessary. Various coherency protocols and specialized bus transfer mechanisms may be employed for this purpose depending on the complexity of the system as well as its requirements. For example, coherence between the cache and the system memory during processor writes may be maintained by employing either a “write-through” or a “write-back” technique. The former technique guarantees consistency between the cache and the system memory by writing the same data to both locations. The latter technique handles coherency by writing only to the cache, and by marking the entry in the cache as being modified. When a modified cache entry is later removed during a cache replacement cycle (or is required by a device other than the processor), the modified data is typically written back to the system memory (and/or provided to the requesting device).
In a multiprocessor shared-memory computer system, separate caches associated with each of the processors may simultaneously store data corresponding to the same memory location. Thus, memory coherency within such systems must typically be handled using somewhat more elaborate and complex schemes. For example, coherency in multiprocessor shared-memory systems may be maintained through employment of either a directory-based protocol or a snooping protocol. In a directory-based protocol, a directory is maintained that indicates which processors have copies of each cache line. This directory is used to limit the processors that must monitor, and possibly respond to, a given request for a cache line. The use of directories reduces snoop traffic and thus allows larger systems to be built. However, the use of directories typically increases the system's latency (which is caused by the directory lookup), as well as the system's hardware complexity and cost.
In a snooping protocol, each processor broadcasts all of its requests for cache lines to all other processors. In many systems, this may be done through a common shared bus. The cache associated with each processor stores along with its address tags coherency information indicating the state of each of its stored lines. Each processor snoops the requests from other processors and responds accordingly by updating its cache tags and/or by providing the data. Thus, each request from another processor may require that a given processor access its own cache's tags to determine if the line exists within the cache, and to update the tag and/or provide the data if necessary. In systems that store cache tags off-chip, the rate at which these cache tags can be accessed may put a limit on the rate at which snoops can be processed and may further limit the bandwidth available for the processor's own cache transactions.
One solution to this problem is to store the cache tags on-chip (on the same chip as the processor), even for cache lines that are stored off-chip. However, this solution suffers from several drawbacks, including the large amount of processor area that must be devoted to maintain these cache tags, the lack of flexibility in changing off-chip cache sizes and organizations, and an increased latency when the data is present in the off-chip cache.
Regardless of whether the cache tags are stored on-chip or off-chip, snoop operations may be necessary to determine whether a remotely requested cache line exists in the cache, whether the cache line state, as reflected in the cache tags, needs to be updated or if data needs to be provided. Performance of such snoop operations may add to latency and processing overhead in computer systems.